On Fri, Dec 27, 2024 at 11:19 AM Steven Rostedt [off-list ref] wrote:

On Thu, 26 Dec 2024 15:01:07 -0800
Linus Torvalds [off-list ref] wrote:

quoted

On Thu, 26 Dec 2024 at 13:49, Steven Rostedt [off-list ref] wrote:

quoted

But then, when the linker removes these functions because they were
overridden, the code does not disappear, leaving the pointers in the
__mcount_loc locations.

This seems entirely unrelated to weak functions, and will be true for
any other "linker removed it" (which can happen for other reasons
too).

So your "fix" seems to be hacking around a symptom.

Yeah, that's why this was just a POC.

quoted

And honestly, the kallsyms argument seems bogus too. The problem with
kallsyms is that it looks up the size the wrong way. Making up new
function names doesn't fix the problem, it - once again - just hacks
around the symptom of doing something wrong.

Christ, kallsyms looking at nm output and going by "next symbol" was
always bogus, but I think that's how the old a.out format worked
originally.

But "nm" literally takes a "-S" argument. We just don't use it.

So I think the fix is literally to just make kallsysms have the size
data. Of course, very annoyingly out /proc/kallsyms file format also
tracks the (legacy) nm output that doesn't have size information.

But I do think that if you hit real problems, you need to fix the
*source* of the issue, not add another ugly hack around things.

So basically the real solution is to fix kallsyms to know about the end
of functions. Peter Zijlstra mentioned that before, but it would take a
bit more work and understanding of kallsyms to fix it properly.

I'm fine not doing the hack and hopefully one day someone will have the
time to fix kallsyms. This was just something I could do quickly,
knowing it is mostly keeping with the status quo and not actually
fixing the root of the issue. I also needed to refresh my ELF parsing
abilities ;-)

I may take a look at kallsyms internals. I have some spare time before
the new year to try and work on things I don't normally get time to
work on.

-- Steve

So, does the kallsyms patch set from Zheng Yejian fix the issues?

It was a long time ago that I reviewed his v1.
I need to take a look if we decide to go with that approach.

-- 
Best Regards
Masahiro Yamada